Skip to content

docs(vision): mlx0330nv campaign — nvfp4 fleet on 0.33.0, and the 12b/31b think-on split - #229

Open
glennneuber wants to merge 1 commit into
mainfrom
docs/vision-campaign-mlx0330-nvfp4
Open

docs(vision): mlx0330nv campaign — nvfp4 fleet on 0.33.0, and the 12b/31b think-on split#229
glennneuber wants to merge 1 commit into
mainfrom
docs/vision-campaign-mlx0330-nvfp4

Conversation

@glennneuber

Copy link
Copy Markdown

The first vsuite campaign on the v0.33.0 fold's build (0.33.0-maxusai-21cfe88e, llama.cpp b10488 / MLX 27fec909), scoped by operator direction to nvfp4 tags only (no GGUF, mxfp8, or bf16). Two files: the campaign doc with both summarizer tables pasted verbatim, and the learnings-log entry for the gemma4:12b think-on runaway.

Findings

  • The fleet is fully functional think-off: all five nvfp4 models (gemma4 12b/26b/31b, qwen3.8:27b, qwen3.6:35b-a3b) converge at the 16384 start rung with scene grounding 6/6 across the board (IoU 0.954–0.999, qwen3.8 best at 0.999 with the only 5/5 name_bbox), perfect serial + invoice extraction. Throughput tracks architecture: 26b / 35b-a3b at 75–90 tok/s (488–581 req/h), dense 31b at 13 tok/s (80 req/h).
  • Think-on splits the gemma4 family. 31b converges at the first rung (27/27 arms, max eval 3,539 of 8,192, quality identical to think-off at 2.4× cost). 12b runs away on the bbox arms under the sanctioned on-policy sampling (card:gemma4+temp0): 25/27 arms capped at 8,192; 20/27 still capped at num_predict 57,344; descoped mid-131072 after ~9.5 h wall for 7 scores. The learnings-log entry is deliberately scope-fenced against the 2026-08-13 finding (which blamed off-policy sampling, on a different build and models) — this is a new observation, not a refutation.
  • The seven 12b arms that did finish are the anchored/pinned ones — consistent with the 2026-08-19 qwen3.6 finding that the anchor bounds reasoning.
  • The 12b think-on cell stays open per ADR 0012 conv 9: its scores file holds 7 finished + 20 capped rows (highest rung 65536) and must not be summarized as results.

Provenance

One host (10.8.0.3, native serve on :11436 via serve-apple-mlx.sh, cold server per cell), one build, campaign tag mlx0330nv1, power mode stamped per cell (the 12b think-off cell predates the powermode 0→2 shift; noted in Limits). Scores untracked on-host per convention; the doc and log entry are the durable record.

🤖 Generated with Claude Code

… 12b/31b think-on split

First vsuite campaign on 0.33.0-maxusai-21cfe88e (b10488 / MLX 27fec909),
nvfp4 tags only by operator direction. All five models fully functional
think-off at the 16384 start rung (scene 6/6 across the fleet, IoU
0.954-0.999). Think-on splits the gemma4 family: 31b converges at rung 1
(27/27 arms, max eval 3,539 of 8,192) while 12b runs away on the bbox arms
under the same card:gemma4+temp0 sampling — 20/27 arms still capped at
num_predict 57,344 — and was descoped mid-ceiling after ~9.5h. The
learnings-log entry records the runaway with its scope fence against the
2026-08-13 off-policy-sampling finding; the 12b think-on cell stays open
per ADR 0012 conv 9.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant